Dataset Overview

Dataset Download

• The datasets can be downloaded from Google Drive and extracted to the datasets/ directory.

Dataset Usage

• The datasets are used for time‑series forecasting experiments, supporting the following models:
  • TSMixer
  • DLinear
  • PatchTST
  • TimesNet

Experiment Examples

Long‑term Forecasting

• Example: use DLinear on the ETTm2 dataset for a forecast length of 96:

python main.py --model DLinear  --data ETTm2 --out_len 96 --in_len 336 --learning_rate 0.001 --batch_size 32 --individual "c"

Zero‑Shot Evaluation

• Example: train DLinear on ETTh1 and perform zero‑shot testing on ETTh2:

python main.py --zero_shot_test True --data ETTh1 --test_data ETTh2 --model DLinear --out_len 96 --individual "c"

M4 Forecasting

• In the M4 dataset, input length and forecast length are specified in datasets/data_loader.py. Train DLinear:

python main_m4.py --model DLinear --data m4 --batch_size 32 --individual "c"

Stock Price Forecasting

• Train DLinear on the Stock dataset with a forecast length of 7:

python main_stock.py --model DLinear --data stock  --out_len 7 --in_len 28 --batch_size 128 --individual "c"

Citation

• If you use this dataset, please cite the related paper:

@article{chen2024similarity,
  title={From Similarity to Superiority: Channel Clustering for Time Series Forecasting},
  author={Chen, Jialin and Lenssen, Jan Eric and Feng, Aosong and Hu, Weihua and Fey, Matthias and Tassiulas, Leandros and Leskovec, Jure and Ying, Rex},
  journal={arXiv preprint arXiv:2404.01340},
  year={2024}
}